IOP PUBLISHING J. Phys. A: Math. Theor.

42 (2009) 042001 (11pp)

JOURNAL OF PHYSICS A: MATHEMATICAL AND THEORETICAL

doi:10.1088/1751-8113/42/4/042001

FAST TRACK COMMUNICATION

Impenetrable barriers in phase space for deterministic thermostats

Gregory S Ezra1 and Stephen Wiggins2
1 Department of Chemistry and Chemical Biology, Baker Laboratory, Cornell University, Ithaca, NY 14853, USA 2 School of Mathematics, University of Bristol, Bristol, BS8 1TW, UK

E-mail: gse1@cornell.edu and stephen.wiggins@mac.com

Received 19 October 2008, in nal form 25 November 2008 Published 17 December 2008 Online at stacks.iop.org/JPhysA/42/042001 Abstract We investigate the relation between the phase-space structures of Hamiltonian and non-Hamiltonian deterministic thermostats. We show that phase-space structures governing reaction dynamics in Hamiltonian systems map to the same type of phase-space structures for the non-Hamiltonian isokinetic equations of motion for the thermostatted Hamiltonian. Our results establish a framework for analyzing thermostat dynamics using concepts and methods developed in reaction rate theory. PACS numbers: 45.05.+x, 45.20.d, 82.20.Db

1. Introduction Deterministic thermostats are widely used to simulate equilibrium physical systems described by ensembles other than microcanonical (constant energy and volume, (E, V )), such as constant temperaturevolume (T , V ) or temperaturepressure (T , p) [15]. Deterministic thermostats are typically obtained by augmenting the phase-space variables of the physical system of interest with a set of additional variables whose role is to alter the standard Hamiltonian system dynamics in such a way that a suitable invariant measure in the system phase space is preserved. In the familiar Nos Hoover (NH) thermostat [6, 7], for example, e the exact dynamics preserves both an extended energy H and a suitable invariant measure, ensuring that, provided the extended system dynamics is effectively ergodic on the timescale of the simulation, the physical system will sample its phase space according to the canonical (constant T) measure. Extended system thermostat dynamics can be either Hamiltonian [6, 811] or nonHamiltonian [7, 1220]. An important motivation for the formulation of Hamiltonian deterministic thermostats such as the Nos Poincar system [11] is the possibility of using e e symplectic integration algorithms to compute trajectories [3, 21, 22].
1751-8113/09/042001+11$30.00 2009 IOP Publishing Ltd Printed in the UK 1

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

In this approach, an extended Hamiltonian is dened for the physical system plus thermostat variables which incorporates a coordinate-dependent time scaling of Poincar e Sundman type [23, 24]. Restricting the dynamics to a xed value (zero) of the extended Hamiltonian results in the system variables sampling their phase space according to, for example, the canonical density [11] (subject to the assumption of ergodicity). The Hamiltonian version of the isokinetic thermostat is described in section 2. A fundamental question concerning deterministic thermostats has to do with the effective ergodicity of the dynamics on the timescale of the simulation. If the dynamics is not effectively ergodic then trajectory simulations will not generate the correct invariant measure [25, 26]. It has long been recognized, for example, that the dynamical system consisting of a single harmonic oscillator degree of freedom coupled to the NH thermostat variable is not ergodic [27]. A large amount of effort has been expended in attempts to design thermostats exhibiting dynamics more ergodic than the basic NH system [3, 4, 28, 29]. The question of ergodicity in thermostats is conceptually closely related to the problem of statistical versus nonstatistical behavior in the (classical) theory of unimolecular reaction rates [3032]. Broadly speaking, in this case one would like to know whether a molecule will behave according to a statistical model such as RRKM theory, or whether it will exhibit signicant deviations from such a theory, ascribable to nonstatistical dynamics [33, 34]. Such nonstatisticality, which can arise from a number of dynamical effects, is analogous to the failure of ergodicity in deterministic thermostats. In recent years, there have been a number of theoretical and computational advances in the application of dynamical systems theory [3537] to study reaction dynamics and phase-space structure in multimode models of molecular systems and to probe the dynamical origins of nonstatistical behavior [38, 39]. The fundamental chemical concept of the transition state, dened as a surface of no return in phase space, has been successfully and rigorously generalized from the well-established 2 degrees of freedom case [40] to systems with N 3 degrees of freedom [39]. Moreover, dynamical indicators exist (determination of reactive phase-space volume, behavior of the reactive ux) to diagnose nonstatistical behavior. Despite their obvious potential relevance for the questions at issue, there has been relatively little work applying the powerful techniques from modern dynamical systems theory, in particular the theory of multidimensional Hamiltonian systems [36, 37], to study the phase-space structure of deterministic thermostats [2, 8, 27, 29, 4144]. There appears to be considerable scope for application of these and other approaches [4547] to the dynamics of deterministic thermostats. In this communication, we begin the development of a novel theoretical framework for the study of thermostat dynamics. Specically, we describe how recently developed methods for the analysis of multimode Hamiltonian systems can be applied to investigate the phase-space structure of the isokinetic thermostat [2]. Although not as widely used as the Nos Hoover thermostat and its many variants, e the non-Hamiltonian version of the isokinetic thermostat has been developed and applied to several problems of chemical interest by Minary et al [48, 49]. In this thermostat, the particle momenta are subject to a nonholonomic constraint that keeps the kinetic energy, hence temperature, constant. The resulting dynamics generates a canonical distribution in conguration space [2]. A Hamiltonian version of the isokinetic thermostat was given by Dettmann [2, 8], and this Hamiltonian formulation (see also [50, 51]) is the point of departure for our investigation. The Hamiltonian formulation of the isokinetic thermostat is presented in section 2. The non-Hamiltonian equations of motion for a Hamiltonian system subject to the isokinetic constraint are shown to correspond to Hamiltonian dynamics at zero energy under an extended
2

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

Hamiltonian whose potential is obtained from the physical potential by exponentiation. The extended Hamiltonian dynamics are therefore nonseparable and potentially chaotic (ergodic), even though the physical Hamiltonian might be separable. For the Hamiltonians we consider that the physical potential exhibits a saddle of index 1, as for the case of a bistable reaction prole coupled to one or more transverse conning modes. The bistable mode can play two distinct roles in the theory: it can either be interpreted as a reaction coordinate of physical interest or as a thermalizing thermostat mode [48]. Essential concepts concerning the phase-space structure of multimode Hamiltonian systems, especially the signicance of normally hyperbolic invariant manifolds (NHIMs) and their role in the phase-space structure and reaction dynamics of multimode molecular systems with index 1 saddle, are briey reviewed in section 3. Our focus in the present study is on saddles of index 1. As mentioned above, the index 1 saddle corresponds to a bistable reaction coordinate coupled to one or more transverse modes, and is a case of fundamental importance for transition state theory. The results obtained in this paper establish the importance of the corresponding structures in the phase space of the isokinetic thermostat. The theory of the phase-space structure in the vicinity of higher index saddles is not as fully developed as the index 1 case, and many open problems remain. In section 4 we show that the extended Hamiltonian dynamical system satises the same conditions satised by the physical Hamiltonian that give rise to the phase-space structures discussed in section 3. We then show that these phase-space structures exist for the non-Hamiltonian isokinetic equations of motion for the thermostatted physical Hamiltonian by an explicit mapping. Section 5 concludes. 2. The physical Hamiltonian, the extended Hamiltonian and non-Hamiltonian isokinetic thermostat We begin with a physical Hamiltonian of the standard form H (q, p) = 1 p2 + 2 (q), (1)

with Hamiltons equations given by q= H = p, p H = q

q (q),

(2a) (2b)

p=

where (q, p) Rn Rn are the physical coordinates and (q) is the potential energy. Following Dettmann and Morriss [2, 8], we construct a Hamiltonian system with the property that trajectories on a xed energy surface of the new Hamiltonian correspond to the trajectories of the physical Hamiltonian (1) which satisfy an isokinetic constraint in the physical coordinates. An extended Hamiltonian K is dened as follows:
K(q, ) = eB HB ,

(3)

where HB is
HB =
1 2

e(B+1) 2 1 e(B1) . 2

(4)

Here, B is an arbitrary parameter, and the relation between the momentum variables p and is specied below. The value chosen for the parameter B denes a particular time scaling via factorization of K; setting B = 1, for example, ensures that HB has q-independent kinetic
3

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

energy. (For simplicity we measure energies in units of kB T , thus keeping the value of T implicit.) The Hamiltonian (3) includes a time scaling factor eB , and Hamiltons equations of motion for K in physical time t are K =e 1 K 1 = q e 2 + e = q 2 2 q=+ (5a) (5b)

and are manifestly B-independent. To show that trajectories of the Hamiltonian system (5) with K = 0 correspond to the trajectories of the physical system (1) satisfying the isokinetic constraint, rst note that the time derivative of HB along trajectories of (5) is given by HB = q HB HB + q
q

(6a) (6b)

=B

e HB .

This implies that trajectories of (5) satisfying HB = 0 at t = 0 satisfy HB = 0 for all t (for arbitrary B). Using (3), this implies that these trajectories are also conned to the surface K = 0 for all the time. The relationship between the Hamiltonian dynamics of (5) on K = 0 and isokinetic trajectories of (1) is made apparent by making the noncanonical transformation of variables q q, e
(q)

(7a) p. (7b)

This coordinate transformation is clearly invertible and is, in fact, a diffeomorphism (as differentiable as ). Applying (7) to HB gives
HB =
1 2

e(B1) (p2 1)

(8)

from which we can immediately conclude that trajectories of the Hamiltonian system (5) with K = HB = 0 automatically satisfy the isokinetic condition p2 = 1, (9) in the physical coordinates (q, p). Substituting relation (7) into (5) we obtain equations of motion for (q, p): q=p p=
2 1 q 2 (p

(10a) + 1) + p(
q

q) =

p,

(10b)

where q p and we have used the constraint p2 = 1. Equations (10) are the isokinetic equations of motion for the thermostatted physical Hamiltonian (1) in physical time t, obtained via Gauss principle of least constraint [2]. By design, the isokinetic dynamics (10) generates a canonical distribution in the coordinates q [2, 8]. Minary et al [48] have shown that the addition of thermalizing degrees of freedom to the physical Hamiltonian (1) can facilitate the attainment of the correct canonical distribution in q-space. If H describes a collection of uncoupled oscillators, the addition
4

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

of a bistable thermalizing degree of freedom renders the Hamiltonian dynamics under K isomorphic to that of a reactive degree of freedom coupled to several bath modes, so that we can obtain useful insights into the thermostat dynamics using methods recently developed for multidimensional Hamiltonian systems. Alternatively, if H describes a reactive mode coupled to bath modes, then the K dynamics is already in an appropriate form for the phase-space analysis described in the following section. 3. Phase-space structures on a xed energy surface Our analysis of thermostat dynamics will be carried out in phase space, using the tools and framework for reaction type dynamics of Hamiltonian systems developed in [38, 39, 5258]. We will show in section 4 that these results apply both to the physical Hamiltonian system (2) and to the extended Hamiltonian system (5). Here we give a brief summary of the setting and relevant results from these references. The starting point for identifying a region of phase space relevant to reaction is to locate an equilibrium point of Hamiltons equations, denoted (q , p ), that is of saddle-center- -center stability type. By this we mean that the matrix associated with the linearization of Hamiltons equations about this equilibrium point has two real eigenvalues of equal magnitude, with one positive and one negative, and n 1 purely imaginary complex conjugate pairs of eigenvalues. We will assume that the purely imaginary eigenvalues satisfy a generic nonresonance condition in the sense that they are independent over the rational numbers (this is discussed in more detail in section 4). We will assume that such an equilibrium point is present in the physical system (2) and we will show that the same type of equilibrium point exists for the extended Hamiltonian system (5) in section 4. However, the discussion in this section applies to any type of Hamiltonian system near the same type of equilibrium point. Without loss of generality we can assume that (q , p ) is located at the origin, and we denote its energy by H (q , p ) h . We will be concerned with geometrical structures in a neighborhood of phase space containing the saddle-center- -center type equilibrium point. We emphasize this fact by denoting the neighborhood by L; this region is to be chosen so that a new set of coordinates can be constructed (the normal form coordinates) in which the Hamiltonian can be expressed (the normal form Hamiltonian) such that it provides an integrable nonlinear approximation to the dynamics which yields phase-space structures to within a given desired accuracy. For h h sufciently small and positive, locally the ((2n 1)-dimensional) energy surface h has the structure of S 2n2 R in the 2n-dimensional phase space. The energy surface h is split locally into two components, reactants (R) and products (P), by a ((2n2)-dimensional) dividing surface (DS(h)) that is diffeomorphic to S 2n2 . The dividing surface that we construct has the following properties: The only way that trajectories can evolve from reactants (R) to products (P) (and vice versa), without leaving the local region L, is through DS(h). In other words, initial conditions on this dividing surface specify all reacting trajectories. The dividing surface is free of local re-crossings; any trajectory which crosses it must leave the neighborhood L before it might possibly cross again. The dividing surface minimizes the (directional) ux. The fundamental phase-space building block that allows the construction of a dividing surface with these properties is a particular normally hyperbolic invariant manifold (NHIM) which, for xed energy h > h , will be denoted by NHIM(h). The NHIM(h) is diffeomorphic to S 2n3 and forms the natural dynamical equator of the dividing surface: the dividing surface
5

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

is split by this equator into (2n 2)-dimensional hemispheres, each diffeomorphic to the open (2n 2) ball, B 2n2 . We will denote these hemispheres by DSf (h) and DSb (h) and call them the forward reactive and backward reactive hemispheres, respectively. DSf (h) is crossed by trajectories representing forward reactions (from reactants to products), while DSb (h) is crossed by trajectories representing backward reactions (from products to reactants). The (2n 3)-dimensional NHIM(h) is an (unstable) invariant subsystem which, in chemistry terminology, corresponds to the energy surface of the activated complex [40, 59]. The NHIM(h) is of saddle stability type, having (2n 2)-dimensional stable and unstable manifolds W s (h) and W u (h) that are diffeomorphic to S 2n3 R. Being of co-dimension one3 with respect to the energy surface, these invariant manifolds act as separatrices, partitioning the energy surface into reacting and nonreacting parts. These phase-space structures can be computed via an algorithmic procedure based on Poincar Birkhoff normalization [38, 39, 52]. This involves developing a new set of e coordinates, the normal form coordinates, (q, p), which are realized through a symplectic coordinate transformation from the original, physical coordinates, T (q, p) = (q, p), (11) which, in a local neighborhood L of the equilibrium point, unfolds the dynamics into a reaction coordinate and bath modes. Expressing H in the new coordinates, (q, p), via HNF (q, p) = H (T 1 (q, p)) (12) gives HNF in a simplied form. The normalization procedure can also be adapted to yield explicit expressions for the coordinate transformations, T (q, p) = (q, p) and T 1 (q, p) = (q, p), between the normal form (NF) coordinates and the original coordinates4 . These coordinate transformations are essential for physical interpretation of the phase-space structures that we construct in normal form coordinates since they allow us to transform these structures back into the original physical coordinates. The nonresonance condition implies that the normal form procedure yields an explicit expression for the normalized Hamiltonian HNF as a function of n local integrals of motion: HNF = HNF (I1 , I2 , . . . , In ). The integral, I1 , corresponds to a reaction coordinate (saddle-type DoF): I1 = q1 p1 . The integrals Ik , for k = 2, . . . , n, correspond to bath modes (center-type DoFs): Ik =
1 2

(13) (14) (15)

2 2 qk + pk .

The integrals provide a natural denition of the term mode that is appropriate in the context of reaction, and their existence is a consequence of the (local) integrability in a neighborhood of the equilibrium point of saddle-center- -center stability type. Moreover, the expression of the normal form Hamiltonian in terms of the integrals provides us with a way to partition the energy between the different modes5 .
3 Briey, the co-dimension of a submanifold is the dimension of the space in which the submanifold exists, minus the dimension of the submanifold. The signicance of a submanifold being co-dimension one is that it is one less dimension than the space in which it exists. Therefore it can divide the space and act as a separatrix, or barrier, to transport. 4 The original coordinates (q, p) had an interpretation as conguration space coordinates and momentum coordinates. The normal form coordinates (q, p), in general, do not have such a physical interpretation since both q and p are nonlinear functions of both q and p. 5 The normal form algorithm that yields all of these results can be applied to realistic molecular Hamiltonians with software available at http://lacms.maths.bris.ac.uk/publications/software/index.html.

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

The n integrals, the normalized Hamiltonian expressed as a function of the integrals and the transformation between the normal form coordinates and the physical coordinates are the key to practically realizing the phase-space structures described at the beginning of this section. The approximate integrability of Hamiltons equations in the reaction region allows a precise and quantitative understanding of all possible trajectories in this region. It also provides a natural construction of an energy dependent reaction coordinate whose properties are determined solely by the Hamiltonian dynamics, as opposed to the need for a priori denitions of possible candidates for reaction coordinates [60]. The n integrals of the motion dened in the neighborhood of the reaction region give rise to further phase-space structures, and therefore constraints on the motion, in addition to those described at the beginning of this section. The common level sets of all the integrals are examples of invariant Lagrangian submanifolds [6163], which have the geometrical structure of two disjoint n-dimensional toroidal cylinders, denoted R Tn1 , i.e. the Cartesian product of a line with n 1 copies of the circle. In the following section we show how all of this phase-space structure exists for thermostatted dynamics of the physical Hamiltonian (1) in physical time t. 4. Microcanonical phase-space structure: Hamiltonian and corresponding non-Hamiltonian thermostatted systems In this section we will show that if the phase-space structure described in section 3 exists for the physical Hamiltonian system (2), it also exists in the phase space of the non-Hamiltonian isokinetic equations of motion (10) corresponding to the thermostatted dynamics of the physical Hamiltonian (1) in physical time t. This is accomplished in three steps by showing: (1) If the physical Hamiltonian system (2) has an equilibrium point at the origin of saddlecenter- -center stability type, then the Hamiltonian system dened by (5) corresponding to the Hamiltonian isokinetic thermostat has an equilibrium point also at the origin of saddle-center- -center stability type. Moreover (and signicantly), we show that the equilibrium points in these two systems satisfy the same nonresonance condition. (2) The energy of the saddle-center- -center type equilibrium point of (5) is negative, but it can be brought sufciently close to zero so that the microcanonical geometrical structures described in section 3 exist on the zero-energy surface of (5). (3) The geometrical structures on the zero-energy surface of (5) map to geometrical structures in the phase space of the non-Hamiltonian thermostatted system corresponding to (10). We begin with step 1. We assume that (2) has an equilibrium point at (q, p) = (q , p ) = (0, 0). From (1), the energy of this equilibrium point is H (0, 0) = (0). The stability of the equilibrium point is determined by the eigenvalues of the derivative of the Hamiltonian vector eld evaluated at the equilibrium point. This is given by the 2n 2n matrix: idnn 0nn , (16) Msys = qq (0) 0nn where 0nn denotes the n n matrix of zeros and idnn denotes the n n identity matrix. We require the equilibrium point to be of saddle-center- -center stability type. This means that the 2n 2n matrix Msys has eigenvalues , ii , i = 2, . . . , n where and i are real. Eigenvalues of Msys are obtained by solving the characteristic equation det(Msys id2n2n ) = 0. From theorem 3 of [64], the block structure of the 2n 2n matrix Msys implies that (17) det(Msys id2n2n ) = det( qq (0) + 2 idnn ) = 0
7

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

so that the 2n eigenvalues are given in terms of , the eigenvalues of the n n Hessian matrix qq (0) associated with the potential as follows: k , k+n = k , k = 1, . . . , n. (18) Therefore, if (q) has a rank-one saddle at q = 0, so that one eigenvalue is strictly negative and the rest are strictly positive, then (q, p) = (0, 0) is a saddle-center- -center type equilibrium point for (2) as described above. We discuss the nonresonance condition in more detail. Suppose 1 < 0 and i > 0. i = 2, . . . , n. Then the nonresonance condition satised by the purely imaginary eigenvalues is given by (m2 , . . . , mn ) (2 , . . . , n ) = 0 for all integer vectors (m2 , . . . , mn ) whose entries are not all zero (where denotes the scalar product). The nonresonance condition is responsible for the existence of the n 1 (local) integrals of motion I2 , . . . , In . Next, we consider the Hamiltonian system (5) corresponding to the Hamiltonian isokinetic thermostat. It is easy to verify that (q, ) = (0, 0) is an equilibrium point for (5) with energy K(0, 0) = 1 e (0) . 2 Proceeding as above, we determine the stability of this equilibrium point by computing the matrix associated with the linearization of (5) at the equilibrium point: Mtherm = 0nn qq (0) 1 e 2
(0)

idnn e+ 0nn

(0)

(19)

The 2n eigenvalues of Mtherm , which we denote as , can be computed by exactly the same type of calculations as above. The resulting eigenvalues are given in terms of the eigenvalues of the potential Hessian as follows: k , k+n = k , 2 k = 1, . . . , n. (20)

Therefore, it is clear that if the potential of the physical Hamiltonian, (q), has a rank-one saddle at q = 0, so that one eigenvalue is strictly negative and the rest are strictly positive, then (q, ) = (0, 0) is a saddle-center- -center type equilibrium point for (5). 1 Moreover, since = 2 it follows by comparing (18) with (20) that if the imaginary parts of the eigenvalues associated with the saddle for the physical Hamiltonian satisfy a nonresonance condition, then they satisfy a nonresonance condition for the saddle associated with the Hamiltonian isokinetic thermostat, i.e (m2 , . . . , mn ) (2 , . . . , n ) = 0 implies that 1 (m2 , . . . , mn ) 2 (2 , . . . , n ) = (m2 , . . . , mn ) (2 , . . . , n ) = 0. Now consider step 2. As we showed above, the saddle-center- -center type equilibrium point (q, ) = (0, 0) of 5 has energy K(0, 0) = 1 e (0) < 0. However, we are only 2 interested in the dynamics on the K = 0 energy surface. The point here is that all of the phase-space structures described in section 3 exist for energies above and sufciently close to the energy of the saddle-center- -center type equilibrium point, and the question is whether or not 1 e (0) < 0 is close enough to zero so that the phase-space structures 2 described in section 3 exist on the K = 0 energy surface. This can easily be arranged by making (0) larger by adding an appropriate constant to (q) or by changing the value of the temperature T. The nal step 3 is to show that the phase-space structure of (5) on K = 0 exists for the isokinetic equations of motion (10) corresponding to the thermostatted dynamics of the physical Hamiltonian (1) in physical time t. Step 3 follows from general results that show that invariant manifolds and their stability properties are preserved under differentiable, invertible (with differentiable inverse) coordinate transformations (i.e. they are preserved under differentiable conjugacies). We emphasize that
8

J. Phys. A: Math. Theor. 42 (2009) 042001

Fast Track Communication

these results are well known and appear in a variety of places throughout the literature, e.g., [65, 66]. These general results allow us to make the following conclusions: Under the map (7) the (2n 1)-dimensional invariant energy surface of the effective Hamiltonian system (5) maps to a (2n 1)-dimensional invariant manifold for the nonHamiltonian isokinetic equations of motion for the thermostatted physical Hamiltonian (1) in physical time t dened by (10). Under the map (7) the (2n 3)-dimensional NHIM, its (2n 2)-dimensional stable and unstable manifolds, the n-dimensional invariant Lagrangian submanifolds and the (2n 2)-dimensional dividing surface map to a (2n 3)-dimensional NHIM, its (2n 2)dimensional stable and unstable manifolds, n-dimensional invariant submanifolds and a (2n 2)-dimensional dividing surface in the (2n 1)-dimensional invariant manifold in the 2n-dimensional phase space of the non-Hamiltonian isokinetic equations of motion for the thermostatted physical Hamiltonian (1) in physical time t dened by (10). 5. Summary and outlook In this paper we have examined the relation between phase-space structures in Hamiltonian and non-Hamiltonian thermostats. In particular, we have established the existence of a mapping between invariant phase-space structures in the phase space of the extended Hamiltonian for the isokinetic thermostat and corresponding structures in the phase space of the non-Hamiltonian Gaussian isokinetic thermostat. Our results establish a conceptual link between the question of thermostat ergodicity and the issue of statisticality in unimolecular isomerization reactions. The existence of normally hyperbolic invariant manifolds in both the physical and extended Hamiltonian phase spaces means that recently developed methods for the analysis of isomerization dynamics can be applied to the thermostat problem. Numerical studies of the isokinetic thermostat based on the ideas presented here are currently in progress. Finally, we note that the approach presented here for the Hamiltonian isokinetic thermostat is in principle applicable to other thermostats. Although the popular Nos Hoover thermostat e [6, 7] and its many variants such as the Nos Hoover chain thermostat [28] are essentially none Hamiltonian dynamical systems, the analysis outlined in the present paper could be applied to Hamiltonian thermostats such as the Nos Poincar system [11]. e e Acknowledgments Some of the original motivation for this work comes from stimulating discussions on phasespace structure in thermostatted systems that arose at the workshop on Metastability and Rare Events in Complex Systems held at the Erwin Schr dinger Institute, Vienna, 1822 o Feb 2008, and organized by Peter Bolhuis, Christoph Dellago and Eric Vanden-Eijnden. SW acknowledges the support of the Ofce of Naval Research Grant No N00014-01-1-0769 and a useful correspondence with Rafael de la Llave concerning the nature of the preservation of phase-space structures under maps. References
[1] Nos S 1991 Prog. Theor. Phys. Suppl. 103 1 e [2] Morriss G P and Dettmann C P 1998 Chaos 8 321 9

J. Phys. A: Math. Theor. 42 (2009) 042001 [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57]

Fast Track Communication

Leimkuhler B and Reich S 2004 Simulating Hamiltonian Dynamics (Cambridge: Cambridge University Press) Hunenberger P 2005 Adv. Polym. Sci. 173 105 Bond S D and Leimkuhler B J 2007 Acta Numer. 16 1 Nos S 1984 J. Chem. Phys. 81 511 e Hoover W G 1985 J. Chem. Phys. 31 1695 Dettmann C P and Morriss G P 1996 Phys. Rev. E 54 2495 Dettmann C P and Morriss G P 1997 Phys. Rev. E 55 3693 Dettmann C P 1999 Phys. Rev. E 60 7576 Bond S D, Leimkuhler B J and Laird B B 1999 J. Comput. Phys. 151 114 Evans D J and Morriss G P 1990 Statistical Mechanics of Nonequilibrium Liquids (New York: Academic) Tuckerman M E, Mundy C J and Martyna G J 1999 Europhys. Lett. 45 149 Tuckerman M E, Liu Y, Ciccotti G and Martyna G J 2001 J. Chem. Phys. 115 1678 Sergi A and Ferrario M 2001 Phys. Rev. E 64 056125 Sergi A 2003 Phys. Rev. E 67 021101 Ezra G S 2004 J. Math. Chem. 35 29 Tarasov V E 2005 J. Phys. A: Math. Gen. 38 2145 Ezra G S 2006 J. Chem. Phys. 125 034104 Sergi A and Giaquinta P V 2007 J. Stat. Mech. P02013 Sanz-Serna J M and Calvo M P 1994 Numerical Hamiltonian Problems (London: Chapman and Hall) Hairer E, Lubich C and Wanner G 2002 Geometric Numerical Integration: Structure Preserving Algorithms for Ordinary Differential Equations (New York: Springer) Bond S D and Leimkuhler B J 1998 Numer. Algorithms 19 55 Benest D and Froeschl C ed 2002 Singularities in Gravitational Systems: Applications to Chaotic Transport e in the Solar System (New York: Springer) Cornfeld I P, Fomin S V and Sinai Y G 1982 Ergodic Theory (New York: Springer) Sturman R, Ottino J M and Wiggins S 2006 The Mathematical Foundations of Mixing (Cambridge: Cambridge University Press) Legoll F, Luskin M and Moeckel R 2007 Arch. Ration. Mech. Anal. 184 449 Martyna G J, Klein M L and Tuckerman M E 1992 J. Chem. Phys. 97 2635 Leimkuhler B J and Sweet C R 2005 SIAM J. Appl. Dyn. Syst. 4 187 Robinson P J and Holbrook K A 1972 Unimolecular Reactions (New York: Wiley) Gilbert R G and Smith S C 1990 Theory of Unimolecular and Recombination Reactions (Oxford: Blackwell)) Baer T and Hase W L 1996 Unimolecular Reaction Dynamics (Oxford: Oxford University Press) DeLeon N and Berne B J 1981 J. Chem. Phys. 75 3495 Carpenter B K 2005 Ann. Rev. Phys. Chem. 56 57 Wiggins S 1990 Physica D 44 471 Wiggins S 1992 Chaotic Transport in Dynamical Systems (Berlin: Springer) Wiggins S 1994 Normally Hyperbolic Invariant Manifolds in Dynamical Systems (Berlin: Springer) Wiggins S, Wiesenfeld L, Jaffe C and Uzer T 2001 Phys. Rev. Lett. 86 5478 Waalkens H, Schubert R and Wiggins S 2008 Nonlinearity 21 R1 Pechukas P 1981 Ann. Rev. Phys. Chem. 32 159 Posch H A, Hoover W G and Vesely F J 1986 Phys. Rev. A 33 4253 Posch H A and Hoover W G 1997 Phys. Rev. E 55 6803 Hoover W G 1998 J. Chem. Phys. 109 4164 Hoover W G, Hoover C G and Isbister D J 2001 Phys. Rev. E 63 026209 Martens C C, Davis M J and Ezra G S 1987 Chem. Phys. Lett. 142 519 Laskar J 1993 Physica D 67 257 Vela-Arevalo L V and Wiggins S 2001 Int. J. Bifurcation Chaos 11 1359 Minary P, Martyna G J and Tuckerman M E 2003 J. Chem. Phys. 118 2510 Minary P, Martyna G J and Tuckerman M E 2003 J. Chem. Phys. 118 2527 Litniewski M 1993 J. Phys. Chem. 97 3842 Morishita T 2003 J. Chem. Phys. 119 7075 Uzer T, Jaffe C, Palacian J, Yanguas P and Wiggins S 2002 Nonlinearity 15 957 Waalkens H, Burbanks A and Wiggins S 2004 J. Phys. A: Math. Gen. 37 L257 Waalkens H and Wiggins S 2004 J. Phys. A: Math. Gen. 37 L435 Waalkens H, Burbanks A and Wiggins S 2004 J. Chem. Phys. 121 6207 Waalkens H, Burbanks A and Wiggins S 2005 Phys. Rev. Lett. 95 084301 Waalkens H, Burbanks A and Wiggins S 2005 J. Phys. A: Math. Gen. 38 L759

10

J. Phys. A: Math. Theor. 42 (2009) 042001 [58] [59] [60] [61] [62] [63] [64] [65] [66]

Fast Track Communication

Schubert R, Waalkens H and Wiggins S 2006 Phys. Rev. Lett. 96 218302 Truhlar D G, Garrett B C and Klippenstein S J 1996 J. Phys. Chem. 100 12771 Heidrich D 1995 The Reaction Path in Chemistry: Current Approaches and Perspectives (Boston: Kluwer) Arnold V I 1978 Mathematical Methods of Classical Mechanics (New York: Springer) Maslov V P and Fedoryuk M V 1981 Semiclassical Approximations in Quantum Mechanics (Boston: Reidel) Littlejohn R G 1992 J. Stat. Phys. 68 7 Silvester J R 2000 Math. Gaz. 84 460 de la Llave R 1992 Commun. Math. Phys. 150 289 Fontich E, de la Llave R and Martn P 2006 Trans. Am. Math. Soc. 358 1317

11